# Systematic analysis of hepatotoxicity: combining literature mining and AI language models

## TL;DR

This paper uses AI and text mining to automatically identify compounds that may cause liver damage, improving toxicity assessment efficiency.

## Contribution

Combines text mining, word embeddings, and large language models to assess hepatotoxicity with high accuracy.

## Key findings

- Text mining achieved an AUC of 0.8 in DILI validation.
- Large language models outperformed text mining with an AUC of 0.85.
- Combining methods improved performance to an AUC of 0.87.

## Abstract

The body of toxicological knowledge and literature is expanding at an accelerating pace. This rapid growth presents significant challenges for researchers, who must stay abreast with latest studies while also synthesizing the vast amount of published information.

Our goal is to automatically identify potential hepatoxicants from over 50,000 compounds using the wealth of scientific publications and knowledge.

We employ and compare three distinct methods for automatic information extraction from unstructured text: (1) text mining (2) word embeddings and (3) large language models. These approaches are combined to calculate a hepatotoxicity score for over 50,000 compounds. We assess the performance of the different methods with a use case on Drug-Induced Liver Injury (DILI).

We evaluated hepatotoxicity for over 50,000 compounds and calculated a hepatotoxicity score for each compound. Our results indicate that text mining is effective for this purpose, achieving an Area Under the Curve (AUC) of 0.8 in DILI validation. Large language models performed even better, with an AUC of 0.85, thanks to their ability to interpret the semantic context accurately. Combining these methods further improved performance, yielding an AUC of 0.87 in DILI validation. All findings are available for download to support further research on toxicity assessment.

We demonstrated that automated text mining is able to successfully assess the toxicity of compounds. A text mining approach seems to be superior to word embeddings. However, the application of a large language model with prompt engineering showed the best performance.